No, it doesn't cost Anthropic $5k per Claude Code user
A recent Forbes article claimed Anthropic's Claude Code Max users could cost the company $5,000 per month, sparking widespread concern about AI inference economics. This story debunks that figure, arguing it confuses retail API prices paid by integrators with Anthropic's far lower actual compute costs, estimated at just 10% of that amount. The revelation challenges popular narratives about the financial unsustainability of AI inference and highlights the significant profit margins frontier AI labs might be making.
The Lowdown
A widely circulated claim suggesting Anthropic's Claude Code Max plan could cost the company $5,000 per power user monthly has been debunked. The author, Martin Alderson, argues this figure misrepresents Anthropic's true compute expenses by conflating them with retail API prices that third-party integrators, such as Cursor, pay.
- The $5,000 estimate originated from a Forbes article discussing Cursor, where it reflects the cost Cursor would incur if it paid Anthropic's retail API prices for power users.
- Alderson uses competitive pricing from OpenRouter for comparable open-weight models (like Qwen 3.5 397B-A17B) to estimate Anthropic's actual compute cost for a power user at approximately $500, roughly 10% of the retail API price.
- He notes that most Claude Code users do not max out their token allocations; Anthropic itself stated that fewer than 5% of subscribers would be affected by weekly caps.
- Based on average usage, Anthropic is likely profitable on a per-user, per-token basis for Claude Code subscribers, with average monthly costs closer to $18 against a $20-200 subscription.
- The author posits that the 'AI inference is a money pit' narrative is misinformation that benefits frontier AI labs by discouraging competition and obscuring their significant profit margins on inference.
Ultimately, while Anthropic faces massive expenses from model training and research, its inference operations are likely profitable, making the $5,000 figure a misattribution of cost from integrator to provider.
The Gossip
Comparative Cost Critiques
Many commenters challenged the article's methodology, particularly the comparison of Claude Opus to open-weight models like Qwen or Kimi. Critics argued that these models might differ significantly in efficiency, quality, or even underlying business models (e.g., Alibaba's infrastructure-as-a-service vs. Anthropic's product focus), making a direct cost comparison flawed. There was extensive discussion on the likely parameter counts and active parameters of Opus 4.6, with some suggesting OpenRouter models are not directly comparable due to unknown quantization or lower performance, while others defended the comparison as a valid baseline for infrastructure costs. The concept of 'opportunity cost' for Anthropic if its compute capacity is fully saturated was also raised as a factor that could push actual costs higher.
User Utilization & Unit Economics
The discussion delved into whether most users truly 'max out' their AI subscriptions. Some asserted that professional users, especially developers, aim to extract maximum value from their paid plans, implying high usage. However, others countered this, drawing parallels to general SaaS trends where most users don't fully utilize their subscriptions. Personal anecdotes from users showed varied usage levels, often falling short of 'maxed out,' suggesting that Anthropic's average profitability per subscriber could be quite high due to low average utilization.
AI Writing Whodunit
A curious side discussion emerged regarding the article's writing style itself. Several commenters suggested that the text exhibited common 'LLM-isms' or stylistic patterns indicative of AI generation or heavy AI editing. They pointed to specific phrases and structural choices. Conversely, others disagreed, stating they found no such tells and that the writing appeared authentically human, leading to a debate on the detectability and prevalence of AI-generated content.